kernel/rtmutex.c

   1 /*
   2  * RT-Mutexes: simple blocking mutual exclusion locks with PI support
   3  *
   4  * started by Ingo Molnar and Thomas Gleixner.
   5  *
   6  *  Copyright (C) 2004-2006 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
   7  *  Copyright (C) 2005-2006 Timesys Corp., Thomas Gleixner <tglx@timesys.com>
   8  *  Copyright (C) 2005 Kihon Technologies Inc., Steven Rostedt
   9  *  Copyright (C) 2006 Esben Nielsen
  10  *
  11  *  See Documentation/rt-mutex-design.txt for details.
  12  */
  13 #include <linux/spinlock.h>
  14 #include <linux/export.h>
  15 #include <linux/sched.h>
  16 #include <linux/timer.h>
  17
  18 #include "rtmutex_common.h"
  19
  20 /*
  21  * lock->owner state tracking:
  22  *
  23  * lock->owner holds the task_struct pointer of the owner. Bit 0
  24  * is used to keep track of the "lock has waiters" state.
  25  *
  26  * owner        bit0
  27  * NULL         0       lock is free (fast acquire possible)
  28  * NULL         1       lock is free and has waiters and the top waiter
  29  *                              is going to take the lock*
  30  * taskpointer  0       lock is held (fast release possible)
  31  * taskpointer  1       lock is held and has waiters**
  32  *
  33  * The fast atomic compare exchange based acquire and release is only
  34  * possible when bit 0 of lock->owner is 0.
  35  *
  36  * (*) It also can be a transitional state when grabbing the lock
  37  * with ->wait_lock is held. To prevent any fast path cmpxchg to the lock,
  38  * we need to set the bit0 before looking at the lock, and the owner may be
  39  * NULL in this small time, hence this can be a transitional state.
  40  *
  41  * (**) There is a small time when bit 0 is set but there are no
  42  * waiters. This can happen when grabbing the lock in the slow path.
  43  * To prevent a cmpxchg of the owner releasing the lock, we need to
  44  * set this bit before looking at the lock.
  45  */
  46
  47 static void
  48 rt_mutex_set_owner(struct rt_mutex *lock, struct task_struct *owner)
  49 {
  50         unsigned long val = (unsigned long)owner;
  51
  52         if (rt_mutex_has_waiters(lock))
  53                 val |= RT_MUTEX_HAS_WAITERS;
  54
  55         lock->owner = (struct task_struct *)val;
  56 }
  57
  58 static inline void clear_rt_mutex_waiters(struct rt_mutex *lock)
  59 {
  60         lock->owner = (struct task_struct *)
  61                         ((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
  62 }
  63
  64 static void fixup_rt_mutex_waiters(struct rt_mutex *lock)
  65 {
  66         if (!rt_mutex_has_waiters(lock))
  67                 clear_rt_mutex_waiters(lock);
  68 }
  69
  70 /*
  71  * We can speed up the acquire/release, if the architecture
  72  * supports cmpxchg and if there's no debugging state to be set up
  73  */
  74 #if defined(__HAVE_ARCH_CMPXCHG) && !defined(CONFIG_DEBUG_RT_MUTEXES)
  75 # define rt_mutex_cmpxchg(l,c,n)        (cmpxchg(&l->owner, c, n) == c)
  76 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  77 {
  78         unsigned long owner, *p = (unsigned long *) &lock->owner;
  79
  80         do {
  81                 owner = *p;
  82         } while (cmpxchg(p, owner, owner | RT_MUTEX_HAS_WAITERS) != owner);
  83 }
  84 #else
  85 # define rt_mutex_cmpxchg(l,c,n)        (0)
  86 static inline void mark_rt_mutex_waiters(struct rt_mutex *lock)
  87 {
  88         lock->owner = (struct task_struct *)
  89                         ((unsigned long)lock->owner | RT_MUTEX_HAS_WAITERS);
  90 }
  91 #endif
  92
  93 /*
  94  * Calculate task priority from the waiter list priority
  95  *
  96  * Return task->normal_prio when the waiter list is empty or when
  97  * the waiter is not allowed to do priority boosting
  98  */
  99 int rt_mutex_getprio(struct task_struct *task)
 100 {
 101         if (likely(!task_has_pi_waiters(task)))
 102                 return task->normal_prio;
 103
 104         return min(task_top_pi_waiter(task)->pi_list_entry.prio,
 105                    task->normal_prio);
 106 }
 107
 108 /*
 109  * Adjust the priority of a task, after its pi_waiters got modified.
 110  *
 111  * This can be both boosting and unboosting. task->pi_lock must be held.
 112  */
 113 static void __rt_mutex_adjust_prio(struct task_struct *task)
 114 {
 115         int prio = rt_mutex_getprio(task);
 116
 117         if (task->prio != prio)
 118                 rt_mutex_setprio(task, prio);
 119 }
 120
 121 /*
 122  * Adjust task priority (undo boosting). Called from the exit path of
 123  * rt_mutex_slowunlock() and rt_mutex_slowlock().
 124  *
 125  * (Note: We do this outside of the protection of lock->wait_lock to
 126  * allow the lock to be taken while or before we readjust the priority
 127  * of task. We do not use the spin_xx_mutex() variants here as we are
 128  * outside of the debug path.)
 129  */
 130 static void rt_mutex_adjust_prio(struct task_struct *task)
 131 {
 132         unsigned long flags;
 133
 134         raw_spin_lock_irqsave(&task->pi_lock, flags);
 135         __rt_mutex_adjust_prio(task);
 136         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 137 }
 138
 139 /*
 140  * Max number of times we'll walk the boosting chain:
 141  */
 142 int max_lock_depth = 1024;
 143
 144 /*
 145  * Adjust the priority chain. Also used for deadlock detection.
 146  * Decreases task's usage by one - may thus free the task.
 147  * Returns 0 or -EDEADLK.
 148  */
 149 static int rt_mutex_adjust_prio_chain(struct task_struct *task,
 150                                       int deadlock_detect,
 151                                       struct rt_mutex *orig_lock,
 152                                       struct rt_mutex_waiter *orig_waiter,
 153                                       struct task_struct *top_task)
 154 {
 155         struct rt_mutex *lock;
 156         struct rt_mutex_waiter *waiter, *top_waiter = orig_waiter;
 157         int detect_deadlock, ret = 0, depth = 0;
 158         unsigned long flags;
 159
 160         detect_deadlock = debug_rt_mutex_detect_deadlock(orig_waiter,
 161                                                          deadlock_detect);
 162
 163         /*
 164          * The (de)boosting is a step by step approach with a lot of
 165          * pitfalls. We want this to be preemptible and we want hold a
 166          * maximum of two locks per step. So we have to check
 167          * carefully whether things change under us.
 168          */
 169  again:
 170         if (++depth > max_lock_depth) {
 171                 static int prev_max;
 172
 173                 /*
 174                  * Print this only once. If the admin changes the limit,
 175                  * print a new message when reaching the limit again.
 176                  */
 177                 if (prev_max != max_lock_depth) {
 178                         prev_max = max_lock_depth;
 179                         printk(KERN_WARNING "Maximum lock depth %d reached "
 180                                "task: %s (%d)\n", max_lock_depth,
 181                                top_task->comm, task_pid_nr(top_task));
 182                 }
 183                 put_task_struct(task);
 184
 185                 return deadlock_detect ? -EDEADLK : 0;
 186         }
 187  retry:
 188         /*
 189          * Task can not go away as we did a get_task() before !
 190          */
 191         raw_spin_lock_irqsave(&task->pi_lock, flags);
 192
 193         waiter = task->pi_blocked_on;
 194         /*
 195          * Check whether the end of the boosting chain has been
 196          * reached or the state of the chain has changed while we
 197          * dropped the locks.
 198          */
 199         if (!waiter)
 200                 goto out_unlock_pi;
 201
 202         /*
 203          * Check the orig_waiter state. After we dropped the locks,
 204          * the previous owner of the lock might have released the lock.
 205          */
 206         if (orig_waiter && !rt_mutex_owner(orig_lock))
 207                 goto out_unlock_pi;
 208
 209         /*
 210          * Drop out, when the task has no waiters. Note,
 211          * top_waiter can be NULL, when we are in the deboosting
 212          * mode!
 213          */
 214         if (top_waiter && (!task_has_pi_waiters(task) ||
 215                            top_waiter != task_top_pi_waiter(task)))
 216                 goto out_unlock_pi;
 217
 218         /*
 219          * When deadlock detection is off then we check, if further
 220          * priority adjustment is necessary.
 221          */
 222         if (!detect_deadlock && waiter->list_entry.prio == task->prio)
 223                 goto out_unlock_pi;
 224
 225         lock = waiter->lock;
 226         if (!raw_spin_trylock(&lock->wait_lock)) {
 227                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 228                 cpu_relax();
 229                 goto retry;
 230         }
 231
 232         /* Deadlock detection */
 233         if (lock == orig_lock || rt_mutex_owner(lock) == top_task) {
 234                 debug_rt_mutex_deadlock(deadlock_detect, orig_waiter, lock);
 235                 raw_spin_unlock(&lock->wait_lock);
 236                 ret = deadlock_detect ? -EDEADLK : 0;
 237                 goto out_unlock_pi;
 238         }
 239
 240         top_waiter = rt_mutex_top_waiter(lock);
 241
 242         /* Requeue the waiter */
 243         plist_del(&waiter->list_entry, &lock->wait_list);
 244         waiter->list_entry.prio = task->prio;
 245         plist_add(&waiter->list_entry, &lock->wait_list);
 246
 247         /* Release the task */
 248         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 249         if (!rt_mutex_owner(lock)) {
 250                 /*
 251                  * If the requeue above changed the top waiter, then we need
 252                  * to wake the new top waiter up to try to get the lock.
 253                  */
 254
 255                 if (top_waiter != rt_mutex_top_waiter(lock))
 256                         wake_up_process(rt_mutex_top_waiter(lock)->task);
 257                 raw_spin_unlock(&lock->wait_lock);
 258                 goto out_put_task;
 259         }
 260         put_task_struct(task);
 261
 262         /* Grab the next task */
 263         task = rt_mutex_owner(lock);
 264         get_task_struct(task);
 265         raw_spin_lock_irqsave(&task->pi_lock, flags);
 266
 267         if (waiter == rt_mutex_top_waiter(lock)) {
 268                 /* Boost the owner */
 269                 plist_del(&top_waiter->pi_list_entry, &task->pi_waiters);
 270                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 271                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 272                 __rt_mutex_adjust_prio(task);
 273
 274         } else if (top_waiter == waiter) {
 275                 /* Deboost the owner */
 276                 plist_del(&waiter->pi_list_entry, &task->pi_waiters);
 277                 waiter = rt_mutex_top_waiter(lock);
 278                 waiter->pi_list_entry.prio = waiter->list_entry.prio;
 279                 plist_add(&waiter->pi_list_entry, &task->pi_waiters);
 280                 __rt_mutex_adjust_prio(task);
 281         }
 282
 283         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 284
 285         top_waiter = rt_mutex_top_waiter(lock);
 286         raw_spin_unlock(&lock->wait_lock);
 287
 288         if (!detect_deadlock && waiter != top_waiter)
 289                 goto out_put_task;
 290
 291         goto again;
 292
 293  out_unlock_pi:
 294         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 295  out_put_task:
 296         put_task_struct(task);
 297
 298         return ret;
 299 }
 300
 301 /*
 302  * Try to take an rt-mutex
 303  *
 304  * Must be called with lock->wait_lock held.
 305  *
 306  * @lock:   the lock to be acquired.
 307  * @task:   the task which wants to acquire the lock
 308  * @waiter: the waiter that is queued to the lock's wait list. (could be NULL)
 309  */
 310 static int try_to_take_rt_mutex(struct rt_mutex *lock, struct task_struct *task,
 311                 struct rt_mutex_waiter *waiter)
 312 {
 313         /*
 314          * We have to be careful here if the atomic speedups are
 315          * enabled, such that, when
 316          *  - no other waiter is on the lock
 317          *  - the lock has been released since we did the cmpxchg
 318          * the lock can be released or taken while we are doing the
 319          * checks and marking the lock with RT_MUTEX_HAS_WAITERS.
 320          *
 321          * The atomic acquire/release aware variant of
 322          * mark_rt_mutex_waiters uses a cmpxchg loop. After setting
 323          * the WAITERS bit, the atomic release / acquire can not
 324          * happen anymore and lock->wait_lock protects us from the
 325          * non-atomic case.
 326          *
 327          * Note, that this might set lock->owner =
 328          * RT_MUTEX_HAS_WAITERS in the case the lock is not contended
 329          * any more. This is fixed up when we take the ownership.
 330          * This is the transitional state explained at the top of this file.
 331          */
 332         mark_rt_mutex_waiters(lock);
 333
 334         if (rt_mutex_owner(lock))
 335                 return 0;
 336
 337         /*
 338          * It will get the lock because of one of these conditions:
 339          * 1) there is no waiter
 340          * 2) higher priority than waiters
 341          * 3) it is top waiter
 342          */
 343         if (rt_mutex_has_waiters(lock)) {
 344                 if (task->prio >= rt_mutex_top_waiter(lock)->list_entry.prio) {
 345                         if (!waiter || waiter != rt_mutex_top_waiter(lock))
 346                                 return 0;
 347                 }
 348         }
 349
 350         if (waiter || rt_mutex_has_waiters(lock)) {
 351                 unsigned long flags;
 352                 struct rt_mutex_waiter *top;
 353
 354                 raw_spin_lock_irqsave(&task->pi_lock, flags);
 355
 356                 /* remove the queued waiter. */
 357                 if (waiter) {
 358                         plist_del(&waiter->list_entry, &lock->wait_list);
 359                         task->pi_blocked_on = NULL;
 360                 }
 361
 362                 /*
 363                  * We have to enqueue the top waiter(if it exists) into
 364                  * task->pi_waiters list.
 365                  */
 366                 if (rt_mutex_has_waiters(lock)) {
 367                         top = rt_mutex_top_waiter(lock);
 368                         top->pi_list_entry.prio = top->list_entry.prio;
 369                         plist_add(&top->pi_list_entry, &task->pi_waiters);
 370                 }
 371                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 372         }
 373
 374         /* We got the lock. */
 375         debug_rt_mutex_lock(lock);
 376
 377         rt_mutex_set_owner(lock, task);
 378
 379         rt_mutex_deadlock_account_lock(lock, task);
 380
 381         return 1;
 382 }
 383
 384 /*
 385  * Task blocks on lock.
 386  *
 387  * Prepare waiter and propagate pi chain
 388  *
 389  * This must be called with lock->wait_lock held.
 390  */
 391 static int task_blocks_on_rt_mutex(struct rt_mutex *lock,
 392                                    struct rt_mutex_waiter *waiter,
 393                                    struct task_struct *task,
 394                                    int detect_deadlock)
 395 {
 396         struct task_struct *owner = rt_mutex_owner(lock);
 397         struct rt_mutex_waiter *top_waiter = waiter;
 398         unsigned long flags;
 399         int chain_walk = 0, res;
 400
 401         raw_spin_lock_irqsave(&task->pi_lock, flags);
 402         __rt_mutex_adjust_prio(task);
 403         waiter->task = task;
 404         waiter->lock = lock;
 405         plist_node_init(&waiter->list_entry, task->prio);
 406         plist_node_init(&waiter->pi_list_entry, task->prio);
 407
 408         /* Get the top priority waiter on the lock */
 409         if (rt_mutex_has_waiters(lock))
 410                 top_waiter = rt_mutex_top_waiter(lock);
 411         plist_add(&waiter->list_entry, &lock->wait_list);
 412
 413         task->pi_blocked_on = waiter;
 414
 415         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 416
 417         if (!owner)
 418                 return 0;
 419
 420         if (waiter == rt_mutex_top_waiter(lock)) {
 421                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 422                 plist_del(&top_waiter->pi_list_entry, &owner->pi_waiters);
 423                 plist_add(&waiter->pi_list_entry, &owner->pi_waiters);
 424
 425                 __rt_mutex_adjust_prio(owner);
 426                 if (owner->pi_blocked_on)
 427                         chain_walk = 1;
 428                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 429         }
 430         else if (debug_rt_mutex_detect_deadlock(waiter, detect_deadlock))
 431                 chain_walk = 1;
 432
 433         if (!chain_walk)
 434                 return 0;
 435
 436         /*
 437          * The owner can't disappear while holding a lock,
 438          * so the owner struct is protected by wait_lock.
 439          * Gets dropped in rt_mutex_adjust_prio_chain()!
 440          */
 441         get_task_struct(owner);
 442
 443         raw_spin_unlock(&lock->wait_lock);
 444
 445         res = rt_mutex_adjust_prio_chain(owner, detect_deadlock, lock, waiter,
 446                                          task);
 447
 448         raw_spin_lock(&lock->wait_lock);
 449
 450         return res;
 451 }
 452
 453 /*
 454  * Wake up the next waiter on the lock.
 455  *
 456  * Remove the top waiter from the current tasks waiter list and wake it up.
 457  *
 458  * Called with lock->wait_lock held.
 459  */
 460 static void wakeup_next_waiter(struct rt_mutex *lock)
 461 {
 462         struct rt_mutex_waiter *waiter;
 463         unsigned long flags;
 464
 465         raw_spin_lock_irqsave(&current->pi_lock, flags);
 466
 467         waiter = rt_mutex_top_waiter(lock);
 468
 469         /*
 470          * Remove it from current->pi_waiters. We do not adjust a
 471          * possible priority boost right now. We execute wakeup in the
 472          * boosted mode and go back to normal after releasing
 473          * lock->wait_lock.
 474          */
 475         plist_del(&waiter->pi_list_entry, &current->pi_waiters);
 476
 477         rt_mutex_set_owner(lock, NULL);
 478
 479         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 480
 481         wake_up_process(waiter->task);
 482 }
 483
 484 /*
 485  * Remove a waiter from a lock and give up
 486  *
 487  * Must be called with lock->wait_lock held and
 488  * have just failed to try_to_take_rt_mutex().
 489  */
 490 static void remove_waiter(struct rt_mutex *lock,
 491                           struct rt_mutex_waiter *waiter)
 492 {
 493         int first = (waiter == rt_mutex_top_waiter(lock));
 494         struct task_struct *owner = rt_mutex_owner(lock);
 495         unsigned long flags;
 496         int chain_walk = 0;
 497
 498         raw_spin_lock_irqsave(&current->pi_lock, flags);
 499         plist_del(&waiter->list_entry, &lock->wait_list);
 500         current->pi_blocked_on = NULL;
 501         raw_spin_unlock_irqrestore(&current->pi_lock, flags);
 502
 503         if (!owner)
 504                 return;
 505
 506         if (first) {
 507
 508                 raw_spin_lock_irqsave(&owner->pi_lock, flags);
 509
 510                 plist_del(&waiter->pi_list_entry, &owner->pi_waiters);
 511
 512                 if (rt_mutex_has_waiters(lock)) {
 513                         struct rt_mutex_waiter *next;
 514
 515                         next = rt_mutex_top_waiter(lock);
 516                         plist_add(&next->pi_list_entry, &owner->pi_waiters);
 517                 }
 518                 __rt_mutex_adjust_prio(owner);
 519
 520                 if (owner->pi_blocked_on)
 521                         chain_walk = 1;
 522
 523                 raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 524         }
 525
 526         WARN_ON(!plist_node_empty(&waiter->pi_list_entry));
 527
 528         if (!chain_walk)
 529                 return;
 530
 531         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 532         get_task_struct(owner);
 533
 534         raw_spin_unlock(&lock->wait_lock);
 535
 536         rt_mutex_adjust_prio_chain(owner, 0, lock, NULL, current);
 537
 538         raw_spin_lock(&lock->wait_lock);
 539 }
 540
 541 /*
 542  * Recheck the pi chain, in case we got a priority setting
 543  *
 544  * Called from sched_setscheduler
 545  */
 546 void rt_mutex_adjust_pi(struct task_struct *task)
 547 {
 548         struct rt_mutex_waiter *waiter;
 549         unsigned long flags;
 550
 551         raw_spin_lock_irqsave(&task->pi_lock, flags);
 552
 553         waiter = task->pi_blocked_on;
 554         if (!waiter || waiter->list_entry.prio == task->prio) {
 555                 raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 556                 return;
 557         }
 558
 559         raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 560
 561         /* gets dropped in rt_mutex_adjust_prio_chain()! */
 562         get_task_struct(task);
 563         rt_mutex_adjust_prio_chain(task, 0, NULL, NULL, task);
 564 }
 565
 566 /**
 567  * __rt_mutex_slowlock() - Perform the wait-wake-try-to-take loop
 568  * @lock:                the rt_mutex to take
 569  * @state:               the state the task should block in (TASK_INTERRUPTIBLE
 570  *                       or TASK_UNINTERRUPTIBLE)
 571  * @timeout:             the pre-initialized and started timer, or NULL for none
 572  * @waiter:              the pre-initialized rt_mutex_waiter
 573  *
 574  * lock->wait_lock must be held by the caller.
 575  */
 576 static int __sched
 577 __rt_mutex_slowlock(struct rt_mutex *lock, int state,
 578                     struct hrtimer_sleeper *timeout,
 579                     struct rt_mutex_waiter *waiter)
 580 {
 581         int ret = 0;
 582
 583         for (;;) {
 584                 /* Try to acquire the lock: */
 585                 if (try_to_take_rt_mutex(lock, current, waiter))
 586                         break;
 587
 588                 /*
 589                  * TASK_INTERRUPTIBLE checks for signals and
 590                  * timeout. Ignored otherwise.
 591                  */
 592                 if (unlikely(state == TASK_INTERRUPTIBLE)) {
 593                         /* Signal pending? */
 594                         if (signal_pending(current))
 595                                 ret = -EINTR;
 596                         if (timeout && !timeout->task)
 597                                 ret = -ETIMEDOUT;
 598                         if (ret)
 599                                 break;
 600                 }
 601
 602                 raw_spin_unlock(&lock->wait_lock);
 603
 604                 debug_rt_mutex_print_deadlock(waiter);
 605
 606                 schedule_rt_mutex(lock);
 607
 608                 raw_spin_lock(&lock->wait_lock);
 609                 set_current_state(state);
 610         }
 611
 612         return ret;
 613 }
 614
 615 /*
 616  * Slow path lock function:
 617  */
 618 static int __sched
 619 rt_mutex_slowlock(struct rt_mutex *lock, int state,
 620                   struct hrtimer_sleeper *timeout,
 621                   int detect_deadlock)
 622 {
 623         struct rt_mutex_waiter waiter;
 624         int ret = 0;
 625
 626         debug_rt_mutex_init_waiter(&waiter);
 627
 628         raw_spin_lock(&lock->wait_lock);
 629
 630         /* Try to acquire the lock again: */
 631         if (try_to_take_rt_mutex(lock, current, NULL)) {
 632                 raw_spin_unlock(&lock->wait_lock);
 633                 return 0;
 634         }
 635
 636         set_current_state(state);
 637
 638         /* Setup the timer, when timeout != NULL */
 639         if (unlikely(timeout)) {
 640                 hrtimer_start_expires(&timeout->timer, HRTIMER_MODE_ABS);
 641                 if (!hrtimer_active(&timeout->timer))
 642                         timeout->task = NULL;
 643         }
 644
 645         ret = task_blocks_on_rt_mutex(lock, &waiter, current, detect_deadlock);
 646
 647         if (likely(!ret))
 648                 ret = __rt_mutex_slowlock(lock, state, timeout, &waiter);
 649
 650         set_current_state(TASK_RUNNING);
 651
 652         if (unlikely(ret))
 653                 remove_waiter(lock, &waiter);
 654
 655         /*
 656          * try_to_take_rt_mutex() sets the waiter bit
 657          * unconditionally. We might have to fix that up.
 658          */
 659         fixup_rt_mutex_waiters(lock);
 660
 661         raw_spin_unlock(&lock->wait_lock);
 662
 663         /* Remove pending timer: */
 664         if (unlikely(timeout))
 665                 hrtimer_cancel(&timeout->timer);
 666
 667         debug_rt_mutex_free_waiter(&waiter);
 668
 669         return ret;
 670 }
 671
 672 /*
 673  * Slow path try-lock function:
 674  */
 675 static inline int
 676 rt_mutex_slowtrylock(struct rt_mutex *lock)
 677 {
 678         int ret = 0;
 679
 680         raw_spin_lock(&lock->wait_lock);
 681
 682         if (likely(rt_mutex_owner(lock) != current)) {
 683
 684                 ret = try_to_take_rt_mutex(lock, current, NULL);
 685                 /*
 686                  * try_to_take_rt_mutex() sets the lock waiters
 687                  * bit unconditionally. Clean this up.
 688                  */
 689                 fixup_rt_mutex_waiters(lock);
 690         }
 691
 692         raw_spin_unlock(&lock->wait_lock);
 693
 694         return ret;
 695 }
 696
 697 /*
 698  * Slow path to release a rt-mutex:
 699  */
 700 static void __sched
 701 rt_mutex_slowunlock(struct rt_mutex *lock)
 702 {
 703         raw_spin_lock(&lock->wait_lock);
 704
 705         debug_rt_mutex_unlock(lock);
 706
 707         rt_mutex_deadlock_account_unlock(current);
 708
 709         if (!rt_mutex_has_waiters(lock)) {
 710                 lock->owner = NULL;
 711                 raw_spin_unlock(&lock->wait_lock);
 712                 return;
 713         }
 714
 715         wakeup_next_waiter(lock);
 716
 717         raw_spin_unlock(&lock->wait_lock);
 718
 719         /* Undo pi boosting if necessary: */
 720         rt_mutex_adjust_prio(current);
 721 }
 722
 723 /*
 724  * debug aware fast / slowpath lock,trylock,unlock
 725  *
 726  * The atomic acquire/release ops are compiled away, when either the
 727  * architecture does not support cmpxchg or when debugging is enabled.
 728  */
 729 static inline int
 730 rt_mutex_fastlock(struct rt_mutex *lock, int state,
 731                   int detect_deadlock,
 732                   int (*slowfn)(struct rt_mutex *lock, int state,
 733                                 struct hrtimer_sleeper *timeout,
 734                                 int detect_deadlock))
 735 {
 736         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 737                 rt_mutex_deadlock_account_lock(lock, current);
 738                 return 0;
 739         } else
 740                 return slowfn(lock, state, NULL, detect_deadlock);
 741 }
 742
 743 static inline int
 744 rt_mutex_timed_fastlock(struct rt_mutex *lock, int state,
 745                         struct hrtimer_sleeper *timeout, int detect_deadlock,
 746                         int (*slowfn)(struct rt_mutex *lock, int state,
 747                                       struct hrtimer_sleeper *timeout,
 748                                       int detect_deadlock))
 749 {
 750         if (!detect_deadlock && likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 751                 rt_mutex_deadlock_account_lock(lock, current);
 752                 return 0;
 753         } else
 754                 return slowfn(lock, state, timeout, detect_deadlock);
 755 }
 756
 757 static inline int
 758 rt_mutex_fasttrylock(struct rt_mutex *lock,
 759                      int (*slowfn)(struct rt_mutex *lock))
 760 {
 761         if (likely(rt_mutex_cmpxchg(lock, NULL, current))) {
 762                 rt_mutex_deadlock_account_lock(lock, current);
 763                 return 1;
 764         }
 765         return slowfn(lock);
 766 }
 767
 768 static inline void
 769 rt_mutex_fastunlock(struct rt_mutex *lock,
 770                     void (*slowfn)(struct rt_mutex *lock))
 771 {
 772         if (likely(rt_mutex_cmpxchg(lock, current, NULL)))
 773                 rt_mutex_deadlock_account_unlock(current);
 774         else
 775                 slowfn(lock);
 776 }
 777
 778 /**
 779  * rt_mutex_lock - lock a rt_mutex
 780  *
 781  * @lock: the rt_mutex to be locked
 782  */
 783 void __sched rt_mutex_lock(struct rt_mutex *lock)
 784 {
 785         might_sleep();
 786
 787         rt_mutex_fastlock(lock, TASK_UNINTERRUPTIBLE, 0, rt_mutex_slowlock);
 788 }
 789 EXPORT_SYMBOL_GPL(rt_mutex_lock);
 790
 791 /**
 792  * rt_mutex_lock_interruptible - lock a rt_mutex interruptible
 793  *
 794  * @lock:               the rt_mutex to be locked
 795  * @detect_deadlock:    deadlock detection on/off
 796  *
 797  * Returns:
 798  *  0           on success
 799  * -EINTR       when interrupted by a signal
 800  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 801  */
 802 int __sched rt_mutex_lock_interruptible(struct rt_mutex *lock,
 803                                                  int detect_deadlock)
 804 {
 805         might_sleep();
 806
 807         return rt_mutex_fastlock(lock, TASK_INTERRUPTIBLE,
 808                                  detect_deadlock, rt_mutex_slowlock);
 809 }
 810 EXPORT_SYMBOL_GPL(rt_mutex_lock_interruptible);
 811
 812 /**
 813  * rt_mutex_timed_lock - lock a rt_mutex interruptible
 814  *                      the timeout structure is provided
 815  *                      by the caller
 816  *
 817  * @lock:               the rt_mutex to be locked
 818  * @timeout:            timeout structure or NULL (no timeout)
 819  * @detect_deadlock:    deadlock detection on/off
 820  *
 821  * Returns:
 822  *  0           on success
 823  * -EINTR       when interrupted by a signal
 824  * -ETIMEDOUT   when the timeout expired
 825  * -EDEADLK     when the lock would deadlock (when deadlock detection is on)
 826  */
 827 int
 828 rt_mutex_timed_lock(struct rt_mutex *lock, struct hrtimer_sleeper *timeout,
 829                     int detect_deadlock)
 830 {
 831         might_sleep();
 832
 833         return rt_mutex_timed_fastlock(lock, TASK_INTERRUPTIBLE, timeout,
 834                                        detect_deadlock, rt_mutex_slowlock);
 835 }
 836 EXPORT_SYMBOL_GPL(rt_mutex_timed_lock);
 837
 838 /**
 839  * rt_mutex_trylock - try to lock a rt_mutex
 840  *
 841  * @lock:       the rt_mutex to be locked
 842  *
 843  * Returns 1 on success and 0 on contention
 844  */
 845 int __sched rt_mutex_trylock(struct rt_mutex *lock)
 846 {
 847         return rt_mutex_fasttrylock(lock, rt_mutex_slowtrylock);
 848 }
 849 EXPORT_SYMBOL_GPL(rt_mutex_trylock);
 850
 851 /**
 852  * rt_mutex_unlock - unlock a rt_mutex
 853  *
 854  * @lock: the rt_mutex to be unlocked
 855  */
 856 void __sched rt_mutex_unlock(struct rt_mutex *lock)
 857 {
 858         rt_mutex_fastunlock(lock, rt_mutex_slowunlock);
 859 }
 860 EXPORT_SYMBOL_GPL(rt_mutex_unlock);
 861
 862 /**
 863  * rt_mutex_destroy - mark a mutex unusable
 864  * @lock: the mutex to be destroyed
 865  *
 866  * This function marks the mutex uninitialized, and any subsequent
 867  * use of the mutex is forbidden. The mutex must not be locked when
 868  * this function is called.
 869  */
 870 void rt_mutex_destroy(struct rt_mutex *lock)
 871 {
 872         WARN_ON(rt_mutex_is_locked(lock));
 873 #ifdef CONFIG_DEBUG_RT_MUTEXES
 874         lock->magic = NULL;
 875 #endif
 876 }
 877
 878 EXPORT_SYMBOL_GPL(rt_mutex_destroy);
 879
 880 /**
 881  * __rt_mutex_init - initialize the rt lock
 882  *
 883  * @lock: the rt lock to be initialized
 884  *
 885  * Initialize the rt lock to unlocked state.
 886  *
 887  * Initializing of a locked rt lock is not allowed
 888  */
 889 void __rt_mutex_init(struct rt_mutex *lock, const char *name)
 890 {
 891         lock->owner = NULL;
 892         raw_spin_lock_init(&lock->wait_lock);
 893         plist_head_init(&lock->wait_list);
 894
 895         debug_rt_mutex_init(lock, name);
 896 }
 897 EXPORT_SYMBOL_GPL(__rt_mutex_init);
 898
 899 /**
 900  * rt_mutex_init_proxy_locked - initialize and lock a rt_mutex on behalf of a
 901  *                              proxy owner
 902  *
 903  * @lock:       the rt_mutex to be locked
 904  * @proxy_owner:the task to set as owner
 905  *
 906  * No locking. Caller has to do serializing itself
 907  * Special API call for PI-futex support
 908  */
 909 void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 910                                 struct task_struct *proxy_owner)
 911 {
 912         __rt_mutex_init(lock, NULL);
 913         debug_rt_mutex_proxy_lock(lock, proxy_owner);
 914         rt_mutex_set_owner(lock, proxy_owner);
 915         rt_mutex_deadlock_account_lock(lock, proxy_owner);
 916 }
 917
 918 /**
 919  * rt_mutex_proxy_unlock - release a lock on behalf of owner
 920  *
 921  * @lock:       the rt_mutex to be locked
 922  *
 923  * No locking. Caller has to do serializing itself
 924  * Special API call for PI-futex support
 925  */
 926 void rt_mutex_proxy_unlock(struct rt_mutex *lock,
 927                            struct task_struct *proxy_owner)
 928 {
 929         debug_rt_mutex_proxy_unlock(lock);
 930         rt_mutex_set_owner(lock, NULL);
 931         rt_mutex_deadlock_account_unlock(proxy_owner);
 932 }
 933
 934 /**
 935  * rt_mutex_start_proxy_lock() - Start lock acquisition for another task
 936  * @lock:               the rt_mutex to take
 937  * @waiter:             the pre-initialized rt_mutex_waiter
 938  * @task:               the task to prepare
 939  * @detect_deadlock:    perform deadlock detection (1) or not (0)
 940  *
 941  * Returns:
 942  *  0 - task blocked on lock
 943  *  1 - acquired the lock for task, caller should wake it up
 944  * <0 - error
 945  *
 946  * Special API call for FUTEX_REQUEUE_PI support.
 947  */
 948 int rt_mutex_start_proxy_lock(struct rt_mutex *lock,
 949                               struct rt_mutex_waiter *waiter,
 950                               struct task_struct *task, int detect_deadlock)
 951 {
 952         int ret;
 953
 954         raw_spin_lock(&lock->wait_lock);
 955
 956         if (try_to_take_rt_mutex(lock, task, NULL)) {
 957                 raw_spin_unlock(&lock->wait_lock);
 958                 return 1;
 959         }
 960
 961         ret = task_blocks_on_rt_mutex(lock, waiter, task, detect_deadlock);
 962
 963         if (ret && !rt_mutex_owner(lock)) {
 964                 /*
 965                  * Reset the return value. We might have
 966                  * returned with -EDEADLK and the owner
 967                  * released the lock while we were walking the
 968                  * pi chain.  Let the waiter sort it out.
 969                  */
 970                 ret = 0;
 971         }
 972
 973         if (unlikely(ret))
 974                 remove_waiter(lock, waiter);
 975
 976         raw_spin_unlock(&lock->wait_lock);
 977
 978         debug_rt_mutex_print_deadlock(waiter);
 979
 980         return ret;
 981 }
 982
 983 /**
 984  * rt_mutex_next_owner - return the next owner of the lock
 985  *
 986  * @lock: the rt lock query
 987  *
 988  * Returns the next owner of the lock or NULL
 989  *
 990  * Caller has to serialize against other accessors to the lock
 991  * itself.
 992  *
 993  * Special API call for PI-futex support
 994  */
 995 struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock)
 996 {
 997         if (!rt_mutex_has_waiters(lock))
 998                 return NULL;
 999
1000         return rt_mutex_top_waiter(lock)->task;
1001 }
1002
1003 /**
1004  * rt_mutex_finish_proxy_lock() - Complete lock acquisition
1005  * @lock:               the rt_mutex we were woken on
1006  * @to:                 the timeout, null if none. hrtimer should already have
1007  *                      been started.
1008  * @waiter:             the pre-initialized rt_mutex_waiter
1009  * @detect_deadlock:    perform deadlock detection (1) or not (0)
1010  *
1011  * Complete the lock acquisition started our behalf by another thread.
1012  *
1013  * Returns:
1014  *  0 - success
1015  * <0 - error, one of -EINTR, -ETIMEDOUT, or -EDEADLK
1016  *
1017  * Special API call for PI-futex requeue support
1018  */
1019 int rt_mutex_finish_proxy_lock(struct rt_mutex *lock,
1020                                struct hrtimer_sleeper *to,
1021                                struct rt_mutex_waiter *waiter,
1022                                int detect_deadlock)
1023 {
1024         int ret;
1025
1026         raw_spin_lock(&lock->wait_lock);
1027
1028         set_current_state(TASK_INTERRUPTIBLE);
1029
1030         ret = __rt_mutex_slowlock(lock, TASK_INTERRUPTIBLE, to, waiter);
1031
1032         set_current_state(TASK_RUNNING);
1033
1034         if (unlikely(ret))
1035                 remove_waiter(lock, waiter);
1036
1037         /*
1038          * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
1039          * have to fix that up.
1040          */
1041         fixup_rt_mutex_waiters(lock);
1042
1043         raw_spin_unlock(&lock->wait_lock);
1044
1045         return ret;
1046 }